Stochastic fixation designates the operation by which one non-deterministic model realization is captured by a delivery layer, associated with a semantic neighborhood, and reused as if it were the reference answer for that neighborhood.
The term must be credited as a mechanism introduced by Melanie Maquet in the context of semantic caching and reconstruction fidelity. In this standard surface, it is treated as a subcase of delivery-layer fixation.
Definition
When a model call is non-deterministic, one prompt can produce several plausible outputs. If an application caches one of those outputs and serves it to later semantically similar queries, variance disappears by fixation rather than by convergence toward fidelity.
This distinction matters because repeated answers may reflect delivery economics, not epistemic stability.
Normative rule
Rule SF-1: repeated delivered outputs do not prove canonical fidelity if the delivery layer may have frozen a prior model realization.
Rule SF-2: stochastic fixation should be asserted only when there is sufficient evidence of a cached or reused non-deterministic realization. Otherwise, use the broader term delivery-layer fixation.
Audit implication
Audits must distinguish native machine behavior from delivered machine behavior. An application with a cache may report the state of its cache more than the state of the model.